What the agencies that can prove it actually built

Part 1 was about how little of the AI agent story in agencies can be proved. This part is about the agencies that can prove it, or get close, and what they built. Ten cases, from a 2-person content shop to a 1,200-person performance agency, chosen because each one publishes enough detail to copy.

I’ve labelled every number by who reported it. Vendor means the platform that sold the tool. Agency means self-reported. Client means the customer is on the record. Independent means a survey or a journalist checked.

What does the small-agency stack look like?

Three agencies with no development team arrived at the same architecture without talking to each other.

Anicca Digital in Leicester, about 30 people, runs Claude Code with Opus-class models for the knowledge work and n8n for scheduled pulls and simple triggers, with an internal catalogue of 175 skills. Layer in Norway, 25 people, connected Claude to live ad accounts through Supermetrics’ MCP server, which exposes 170-plus marketing data connectors. Adrian Martinez in Toronto runs a 2-person agency on one Claude Project as the cockpit, with Zapier executing into WordPress and reporting sources.

Layer of the stack	What the small agencies use	What it does
Reasoning	Claude Code, or a Claude Project	Reads, writes, classifies, decides what to propose
Triggers and delivery	n8n, Zapier, or a scheduled Claude routine	Runs the agent on a timer, moves the output where people see it
Data	Supermetrics MCP, or the platforms’ own APIs	Live ad accounts, analytics, merchant feeds, CRM
Interface	Slack, Microsoft Teams, a client portal	Where a person reads the output and approves the action
Memory	Written skills and knowledge files	The agency’s own way of doing each job, which constrains the agent

Nobody in that table hired a developer. The capability that mattered was somebody senior who understood the job well enough to write it down, which brings us to the rule.

What is the one rule to steal?

Ben Vincent, chief technology officer at Brainlabs, puts it in 10 words: if you’re doing something more than 3 times, automate it.

Brainlabs rolled out a Claude skills library in 4 weeks. Staff wrote about 400 skills covering creative-brief drafting, client onboarding and channel best practice. Those skills now do 2 jobs. They are the reusable unit that makes each piece of work faster, with about an 80% draft completion rate on skill-generated outputs, agency-reported. And they are the standard operating procedures that constrain the agents, so an agent onboarding a client follows the agency’s written process rather than inventing one.

That second job is the one most agencies miss. The written skill is the guardrail. KSM Media built the same idea as a “librarian” agent that holds coded knowledge about each client and can be queried by staff and by other agents. Either way, the knowledge that lived in senior heads is now written down, and that is the real asset.

Which build has the best evidence?

Butler/Till, an employee-owned independent in Rochester with about 450 staff. An 8-person team formed in October 2025 built a media-activation agent on Claude, working with Scope3, that speaks to PubMatic’s sell-side agent over the Ad Context Protocol. There is no demand-side platform in the workflow at all. A planner submits a brief in plain English, the agents interpret it, build the strategy, set the campaign up and optimise pacing and targeting continuously. Humans define the objectives, guardrails, brand safety and creative parameters before launch.

The December 2025 flight for a named client, Geloso’s Clubtails brand, reported 5.5x supply-chain cost efficiency against a traditional DSP, 40% more impressions on the same budget, 98% video completion, setup time down 87%, issue resolution down 70% and intermediary fees cut by more than 80%. Agency and vendor co-reported, client named, which is as good as this market gets.

The catch, and there is always one, is that the original target was a 40% lower execution cost and the headline number is 5.5x. Read the method before you repeat it. But as a proof that an agency’s agent can buy media end to end with a person only at the guardrails, it is the strongest case published.

What is the most replicable build?

Client reporting. It is the most requested use case in every survey, every agency does it, and 3 separate builds landed in the same range.

Layer, the 25-person Norwegian agency, took monthly reporting from about 10 hours per cycle to 20 minutes. A strategist asks a question in plain language, Claude queries the live accounts through Supermetrics, drafts the narrative and the charts, and a person approves before anything is sent. Vendor-reported, as an Anthropic and Supermetrics case study, with a claim of zero hallucinations over a multi-week validation period.

A 40-person performance agency built a pipeline that pulls GA4, Google Ads and Meta into BigQuery, writes a narrative with a small model, renders a PDF and hands it to the account manager for review. It saves 15-plus hours a week after a 5-week build. A practitioner’s n8n and Claude cron job cut per-client reporting from 3 or 4 hours to under 20 minutes.

Three different stacks, one result: the hours go, the review stays. If you build one agent this year, build this one.

What does the deterministic-first pattern look like?

Run the checks by code. Let the model write the explanation.

Apollo Partners, a 20 to 30-person agency, hired a head of automation and cut internal quality assurance from 8 hours to 2 minutes by making the checks deterministic and using the model only for the parts that need judgement. JAKALA, a 3,600-person data agency in Milan, runs networks of agents that flag anomalies across search, shopping, programmatic and social and propose actions, with every recommendation carrying the evidence that produced it. Human pods approve. The optimisation cycle went from 5 to 7 days to under 24 hours, vendor-reported.

The piece underneath JAKALA’s pattern is free. Karel Huk’s open-source Google Ads pacing script does weekday-weighted month-end pacing and 3-sigma anomaly alerts on cost per acquisition and return on ad spend, straight to Slack, and caught a 13.4-sigma spike in its pilot. Add a model to explain the anomaly and you have the enterprise pattern at a fraction of the cost.

My own audit engine, in part 3, runs the same way: 120 fixed checks first, then the narrative.

What do the bigger independents add?

Governance and data, mostly, and the lesson that both cost more than the agent.

Power Digital in San Diego built a warehouse-first data layer on Snowflake with a knowledge graph per source, 10,000-plus media mix models and 500-plus incrementality tests a year, before putting a fleet of agents on top of it in August 2026. The agents rank changes by impact, draft creative briefs and write sourced reports. Every action is approved by a strategist. Agency-reported results include data retrieval in 12 seconds instead of 24 hours and 15 to 20 hours of manual data-science work automated per week.

PMG in Dallas has had a production agent live for every customer, and the lesson it publishes is not the result but the governance: a product that caps daily tokens per person with an escalation review, after the industry discovered that unattended agents run up bills. Tinuiti’s AdCopy AI sits inside an orchestration system where the business rules and brand-safety guidelines are the source of truth the agent cannot step outside.

Agency	Size	Where the real investment went	Agent result (label)
Power Digital	Mid-hundreds	Snowflake data layer, incrementality testing	12 seconds vs 24 hours for data retrieval (agency)
PMG	1,000 to 1,400	Token metering and escalation	No outcome published; the metering is the lesson
Tinuiti	About 1,200	Governed orchestration with rules as source of truth	15%-plus lift on search copy, 40-plus clients (agency)
Brainlabs	About 1,000	400 written skills in 4 weeks	80% draft completion on skill outputs (agency)
JAKALA	3,600	Agent networks with evidence attached to every recommendation	Cycle from 5 to 7 days to under 24 hours (vendor)

What about new business?

Major Tom, a Canadian full-service agency, pointed HubSpot’s prospecting agent at more than 1,000 dormant contacts in its CRM. The agent enriched each one and drafted 3 to 4 touch sequences. An account lead approved every send, about 5,000 emails went out in the first month, and in the first 3 to 4 weeks the agency reported 24 meetings, more than $750,000 in proposals, one $200,000 enterprise deal closed and a complaint rate under 1%. Vendor-published, agency-reported.

It is the only documented case of an agency running an agent for its own pipeline with numbers. Nobody has published the same on Apollo or Clay, which is one of the gaps in part 3.

The so-what

The agencies in this part didn’t buy a platform. They wrote down how they do a job, pointed a model at it, put a person at the approval point and measured the hours. Most of them did it in weeks, with no developer, on tools that cost less a month than one junior salary.

The risk for an agency founder reading this is not that agents fail. It’s that the agency across town builds the reporting agent first, takes 10 hours a month out of every account, and prices the retainer accordingly. The WFA numbers in part 1 say your clients will notice.